mim_cmt_finetune.py

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.checkpoint as checkpoint

import math
import logging
from functools import partial
from collections import OrderedDict

from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
from timm.models.layers import DropPath, to_2tuple, trunc_normal_
from timm.models.registry import register_model

_logger = logging.getLogger(__name__)


def _cfg(url='', **kwargs):
    return {
        'url': url,
        'num_classes': 1000, 'input_size': (3, 224, 224), 'pool_size': None,
        'crop_pct': .9, 'interpolation': 'bicubic',
        'mean': IMAGENET_DEFAULT_MEAN, 'std': IMAGENET_DEFAULT_STD,
        'first_conv': 'patch_embed.proj', 'classifier': 'head',
        **kwargs
    }

# A memory-efficient implementation of Swish function
class SwishImplementation(torch.autograd.Function):
    @staticmethod
    def forward(ctx, i):
        result = i * torch.sigmoid(i)
        ctx.save_for_backward(i)
        return result

    @staticmethod
    def backward(ctx, grad_output):
        i = ctx.saved_tensors[0]
        sigmoid_i = torch.sigmoid(i)
        return grad_output * (sigmoid_i * (1 + i * (1 - sigmoid_i)))


class MemoryEfficientSwish(nn.Module):
    def forward(self, x):
        return SwishImplementation.apply(x)


class Mlp(nn.Module):
    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
        super().__init__()
        out_features = out_features or in_features
        hidden_features = hidden_features or in_features
        self.conv1 = nn.Sequential(
            nn.Conv2d(in_features, hidden_features, 1, 1, 0, bias=True),
            nn.GELU(),
            nn.BatchNorm2d(hidden_features, eps=1e-5),
        )
        self.proj = nn.Conv2d(hidden_features, hidden_features, 3, 1, 1, groups=hidden_features)
        self.proj_act = nn.GELU()
        self.proj_bn = nn.BatchNorm2d(hidden_features, eps=1e-5)
        self.conv2 = nn.Sequential(
            nn.Conv2d(hidden_features, out_features, 1, 1, 0, bias=True),
            nn.BatchNorm2d(out_features, eps=1e-5),
        )
        self.drop = nn.Dropout(drop)

    def forward(self, x, H, W):
        B, N, C = x.shape
        x = x.permute(0, 2, 1).reshape(B, C, H, W)
        x = self.conv1(x)
        x = self.drop(x)
        x = self.proj(x) + x
        x = self.proj_act(x)
        x = self.proj_bn(x)
        x = self.conv2(x)
        x = x.flatten(2).permute(0, 2, 1)
        x = self.drop(x)
        return x


class Attention(nn.Module):
    def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, 
                 attn_drop=0., proj_drop=0., qk_ratio=1, sr_ratio=1):
        super().__init__()
        self.num_heads = num_heads
        head_dim = dim // num_heads
        self.scale = qk_scale or head_dim ** -0.5
        self.qk_dim = dim // qk_ratio

        self.q = nn.Linear(dim, self.qk_dim, bias=qkv_bias)
        self.k = nn.Linear(dim, self.qk_dim, bias=qkv_bias)
        self.v = nn.Linear(dim, dim, bias=qkv_bias)
        self.attn_drop = nn.Dropout(attn_drop)
        self.proj = nn.Linear(dim, dim)
        self.proj_drop = nn.Dropout(proj_drop)
        
        self.sr_ratio = sr_ratio
        # Exactly same as PVTv1
        if self.sr_ratio > 1:
            self.sr = nn.Sequential(
                nn.Conv2d(dim, dim, kernel_size=sr_ratio, stride=sr_ratio, groups=dim, bias=True),
                nn.BatchNorm2d(dim, eps=1e-5),
            )

    def forward(self, x, H, W, relative_pos):
        B, N, C = x.shape
        q = self.q(x).reshape(B, N, self.num_heads, self.qk_dim // self.num_heads).permute(0, 2, 1, 3)
        
        if self.sr_ratio > 1:
            x_ = x.permute(0, 2, 1).reshape(B, C, H, W)
            x_ = self.sr(x_).reshape(B, C, -1).permute(0, 2, 1)
            k = self.k(x_).reshape(B, -1, self.num_heads, self.qk_dim // self.num_heads).permute(0, 2, 1, 3)
            v = self.v(x_).reshape(B, -1, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
        else:
            k = self.k(x).reshape(B, N, self.num_heads, self.qk_dim // self.num_heads).permute(0, 2, 1, 3)
            v = self.v(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)

        attn = (q @ k.transpose(-2, -1)) * self.scale + relative_pos
        attn = attn.softmax(dim=-1)
        attn = self.attn_drop(attn)
        x = (attn @ v).transpose(1, 2).reshape(B, N, C)
        x = self.proj(x)
        x = self.proj_drop(x)
        return x


class Block(nn.Module):
    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
                 drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, qk_ratio=1, sr_ratio=1):
        super().__init__()
        self.norm1 = norm_layer(dim)
        self.attn = Attention(
            dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, 
            attn_drop=attn_drop, proj_drop=drop, qk_ratio=qk_ratio, sr_ratio=sr_ratio)
        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
        self.norm2 = norm_layer(dim)
        mlp_hidden_dim = int(dim * mlp_ratio)
        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
        self.proj = nn.Conv2d(dim, dim, 3, 1, 1, groups=dim)
        
    def forward(self, x, H, W, relative_pos):
        B, N, C = x.shape
        cnn_feat = x.permute(0, 2, 1).reshape(B, C, H, W)
        x = self.proj(cnn_feat) + cnn_feat
        x = x.flatten(2).permute(0, 2, 1)
        x = x + self.drop_path(self.attn(self.norm1(x), H, W, relative_pos))
        x = x + self.drop_path(self.mlp(self.norm2(x), H, W))
        return x


class PatchEmbed(nn.Module):
    """ Image to Patch Embedding
    """
    def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768):
        super().__init__()
        img_size = to_2tuple(img_size)
        patch_size = to_2tuple(patch_size)
        num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0])
        
        assert img_size[0] % patch_size[0] == 0 and img_size[1] % patch_size[1] == 0, \
            f"img_size {img_size} should be divided by patch_size {patch_size}."
        
        self.img_size = img_size
        self.patch_size = patch_size
        self.num_patches = num_patches

        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
        self.norm = nn.LayerNorm(embed_dim)

    def forward(self, x):
        B, C, H, W = x.shape
        # FIXME look at relaxing size constraints
        assert H == self.img_size[0] and W == self.img_size[1], \
            f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
        x = self.proj(x).flatten(2).transpose(1, 2)
        x = self.norm(x)
        
        H, W = H // self.patch_size[0], W // self.patch_size[1]
        return x, (H, W)


class CMT(nn.Module):
    def __init__(self, img_size=224, in_chans=3, num_classes=1000, embed_dims=[46,92,184,368], stem_channel=16, fc_dim=1280,
                 num_heads=[1,2,4,8], mlp_ratios=[3.6,3.6,3.6,3.6], qkv_bias=True, qk_scale=None, representation_size=None,
                 drop_rate=0., attn_drop_rate=0., drop_path_rate=0., hybrid_backbone=None, norm_layer=None,
                 depths=[2,2,10,2], qk_ratio=1, sr_ratios=[8,4,2,1], dp=0.1, **kwargs):
        super().__init__()
        self.num_classes = num_classes
        self.num_features = self.embed_dim = embed_dims[-1]
        norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6)
        self.depths = depths
        
        self.stem_conv1 = nn.Conv2d(3, stem_channel, kernel_size=3, stride=2, padding=1, bias=True)
        self.stem_relu1 = nn.GELU()
        self.stem_norm1 = nn.BatchNorm2d(stem_channel, eps=1e-5)
        
        self.stem_conv2 = nn.Conv2d(stem_channel, stem_channel, kernel_size=3, stride=1, padding=1, bias=True)
        self.stem_relu2 = nn.GELU()
        self.stem_norm2 = nn.BatchNorm2d(stem_channel, eps=1e-5)
        
        self.stem_conv3 = nn.Conv2d(stem_channel, stem_channel, kernel_size=3, stride=1, padding=1, bias=True)
        self.stem_relu3 = nn.GELU()
        self.stem_norm3 = nn.BatchNorm2d(stem_channel, eps=1e-5)

        self.patch_embed_a = PatchEmbed(
            img_size=img_size//2, patch_size=2, in_chans=stem_channel, embed_dim=embed_dims[0])
        self.patch_embed_b = PatchEmbed(
            img_size=img_size//4, patch_size=2, in_chans=embed_dims[0], embed_dim=embed_dims[1])
        self.patch_embed_c = PatchEmbed(
            img_size=img_size//8, patch_size=2, in_chans=embed_dims[1], embed_dim=embed_dims[2])
        self.patch_embed_d = PatchEmbed(
            img_size=img_size//16, patch_size=2, in_chans=embed_dims[2], embed_dim=embed_dims[3])

        self.relative_pos_a = nn.Parameter(torch.randn(
            num_heads[0], self.patch_embed_a.num_patches, self.patch_embed_a.num_patches//sr_ratios[0]//sr_ratios[0]))
        self.relative_pos_b = nn.Parameter(torch.randn(
            num_heads[1], self.patch_embed_b.num_patches, self.patch_embed_b.num_patches//sr_ratios[1]//sr_ratios[1]))
        self.relative_pos_c = nn.Parameter(torch.randn(
            num_heads[2], self.patch_embed_c.num_patches, self.patch_embed_c.num_patches//sr_ratios[2]//sr_ratios[2]))
        self.relative_pos_d = nn.Parameter(torch.randn(
            num_heads[3], self.patch_embed_d.num_patches, self.patch_embed_d.num_patches//sr_ratios[3]//sr_ratios[3]))
        
        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]  # stochastic depth decay rule
        _logger.info('drop path rate: %s', drop_path_rate)
        cur = 0
        self.blocks_a = nn.ModuleList([
            Block(
                dim=embed_dims[0], num_heads=num_heads[0], mlp_ratio=mlp_ratios[0], qkv_bias=qkv_bias,
                qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur+i],
                norm_layer=norm_layer, qk_ratio=qk_ratio, sr_ratio=sr_ratios[0])
            for i in range(depths[0])])
        cur += depths[0]
        self.blocks_b = nn.ModuleList([
            Block(
                dim=embed_dims[1], num_heads=num_heads[1], mlp_ratio=mlp_ratios[1], qkv_bias=qkv_bias,
                qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur+i],
                norm_layer=norm_layer, qk_ratio=qk_ratio, sr_ratio=sr_ratios[1])
            for i in range(depths[1])])
        cur += depths[1]
        self.blocks_c = nn.ModuleList([
            Block(
                dim=embed_dims[2], num_heads=num_heads[2], mlp_ratio=mlp_ratios[2], qkv_bias=qkv_bias,
                qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur+i],
                norm_layer=norm_layer, qk_ratio=qk_ratio, sr_ratio=sr_ratios[2])
            for i in range(depths[2])])
        cur += depths[2]
        self.blocks_d = nn.ModuleList([
            Block(
                dim=embed_dims[3], num_heads=num_heads[3], mlp_ratio=mlp_ratios[3], qkv_bias=qkv_bias,
                qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur+i],
                norm_layer=norm_layer, qk_ratio=qk_ratio, sr_ratio=sr_ratios[3])
            for i in range(depths[3])])

        # Representation layer
        if representation_size:
            self.num_features = representation_size
            self.pre_logits = nn.Sequential(OrderedDict([
                ('fc', nn.Linear(embed_dim, representation_size)),
                ('act', nn.Tanh())
            ]))
        else:
            self.pre_logits = nn.Identity()

        # Classifier head
        self._fc = nn.Conv2d(embed_dims[-1], fc_dim, kernel_size=1)
        self._bn = nn.BatchNorm2d(fc_dim, eps=1e-5)
        self._swish = MemoryEfficientSwish()
        self._avg_pooling = nn.AdaptiveAvgPool2d(1)
        self._drop = nn.Dropout(dp)
        self.head = nn.Linear(fc_dim, num_classes) if num_classes > 0 else nn.Identity()
        self.apply(self._init_weights)

    def _init_weights(self, m):
        if isinstance(m, nn.Linear):
            trunc_normal_(m.weight, std=.02)
            if isinstance(m, nn.Linear) and m.bias is not None:
                nn.init.constant_(m.bias, 0)
        elif isinstance(m, nn.Conv2d):
            nn.init.kaiming_normal_(m.weight, mode='fan_out')
            if isinstance(m, nn.Conv2d) and m.bias is not None:
                nn.init.constant_(m.bias, 0)
        elif isinstance(m, nn.LayerNorm):
            nn.init.constant_(m.bias, 0)
            nn.init.constant_(m.weight, 1.0)
        elif isinstance(m, nn.BatchNorm2d):
            nn.init.constant_(m.bias, 0)
            nn.init.constant_(m.weight, 1.0)
            
    def update_temperature(self):
        for m in self.modules():
            if isinstance(m, Attention):
                m.update_temperature()

    @torch.jit.ignore
    def no_weight_decay(self):
        return {'pos_embed', 'cls_token'}

    def get_classifier(self):
        return self.head

    def reset_classifier(self, num_classes, global_pool=''):
        self.num_classes = num_classes
        self.head = nn.Linear(self.embed_dims[-1], num_classes) if num_classes > 0 else nn.Identity()
    
    def get_num_layers(self):
        return self.depths

    def forward_features(self, x):
        B = x.shape[0]
        x = self.stem_conv1(x)
        x = self.stem_relu1(x)
        x = self.stem_norm1(x)
        
        x = self.stem_conv2(x)
        x = self.stem_relu2(x)
        x = self.stem_norm2(x)
        
        x = self.stem_conv3(x)
        x = self.stem_relu3(x)
        x = self.stem_norm3(x)
        
        x, (H, W) = self.patch_embed_a(x)
        for i, blk in enumerate(self.blocks_a):
            x = blk(x, H, W, self.relative_pos_a)
            
        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
        x, (H, W) = self.patch_embed_b(x)
        for i, blk in enumerate(self.blocks_b):
            x = blk(x, H, W, self.relative_pos_b)
            
        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
        x, (H, W) = self.patch_embed_c(x)
        for i, blk in enumerate(self.blocks_c):
            x = blk(x, H, W, self.relative_pos_c)
            
        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
        x, (H, W) = self.patch_embed_d(x)
        for i, blk in enumerate(self.blocks_d):
            x = blk(x, H, W, self.relative_pos_d)

        B, N, C = x.shape
        x = self._fc(x.permute(0, 2, 1).reshape(B, C, H, W))
        x = self._bn(x)
        x = self._swish(x)
        x = self._avg_pooling(x).flatten(start_dim=1)
        x = self._drop(x)
        x = self.pre_logits(x)
        return x

    def forward(self, x):
        x = self.forward_features(x)
        x = self.head(x)
        return x


def resize_pos_embed(posemb, posemb_new):
    # Rescale the grid of position embeddings when loading from state_dict. Adapted from
    # https://github.com/google-research/vision_transformer/blob/00883dd691c63a6830751563748663526e811cee/vit_jax/checkpoint.py#L224
    _logger.info('Resized position embedding: %s to %s', posemb.shape, posemb_new.shape)
    ntok_new = posemb_new.shape[1]
    if True:
        posemb_tok, posemb_grid = posemb[:, :1], posemb[0, 1:]
        ntok_new -= 1
    else:
        posemb_tok, posemb_grid = posemb[:, :0], posemb[0]
    gs_old = int(math.sqrt(len(posemb_grid)))
    gs_new = int(math.sqrt(ntok_new))
    _logger.info('Position embedding grid-size from %s to %s', gs_old, gs_new)
    posemb_grid = posemb_grid.reshape(1, gs_old, gs_old, -1).permute(0, 3, 1, 2)
    posemb_grid = F.interpolate(posemb_grid, size=(gs_new, gs_new), mode='bilinear')
    posemb_grid = posemb_grid.permute(0, 2, 3, 1).reshape(1, gs_new * gs_new, -1)
    posemb = torch.cat([posemb_tok, posemb_grid], dim=1)
    return posemb


def checkpoint_filter_fn(state_dict, model):
    """ convert patch embedding weight from manual patchify + linear proj to conv"""
    out_dict = {}
    if 'model' in state_dict:
        # For deit models
        state_dict = state_dict['model']
    for k, v in state_dict.items():
        if 'patch_embed.proj.weight' in k and len(v.shape) < 4:
            # For old models that I trained prior to conv based patchification
            O, I, H, W = model.patch_embed.proj.weight.shape
            v = v.reshape(O, -1, H, W)
        elif k == 'pos_embed' and v.shape != model.pos_embed.shape:
            # To resize pos embedding when using model at different size from pretrained weights
            v = resize_pos_embed(v, model.pos_embed)
        out_dict[k] = v
    return out_dict


def _create_cmt_model(pretrained=False, distilled=False, **kwargs):
    default_cfg = _cfg()
    default_num_classes = default_cfg['num_classes']
    default_img_size = default_cfg['input_size'][-1]

    num_classes = kwargs.pop('num_classes', default_num_classes)
    img_size = kwargs.pop('img_size', default_img_size)
    repr_size = kwargs.pop('representation_size', None)
    if repr_size is not None and num_classes != default_num_classes:
        # Remove representation layer if fine-tuning. This may not always be the desired action,
        # but I feel better than doing nothing by default for fine-tuning. Perhaps a better interface?
        _logger.warning("Removing representation layer for fine-tuning.")
        repr_size = None

    model = CMT(img_size=img_size, num_classes=num_classes, representation_size=repr_size, **kwargs)
    model.default_cfg = default_cfg

    if pretrained:
        load_pretrained(
            model, num_classes=num_classes, in_chans=kwargs.get('in_chans', 3),
            filter_fn=partial(checkpoint_filter_fn, model=model))
    return model


@register_model
def cmt_small(pretrained=False, **kwargs):
    model_kwargs = dict(
        qkv_bias=True, embed_dims=[64,128,256,512], stem_channel=32, num_heads=[1,2,4,8],
        depths=[3,3,16,3], mlp_ratios=[4,4,4,4], qk_ratio=1, sr_ratios=[8,4,2,1], **kwargs)
    model = _create_cmt_model(pretrained=pretrained, **model_kwargs)
    return model